The start-up company Q.ANT has announced the market launch of a photonic Native Processing Unit (NPU). It is based on the company's own computing architecture LENA (Light Empowered Native Arithmetics) and builds on the industry standard PCI Express. The NPU can perform complex, non-linear calculations natively using light instead of electrons.
The technology promises significantly higher energy efficiency and improved computing speed compared to conventional CMOS technology for applications such as AI inference, machine learning and physics simulations. The company cites the solution of computational problems in number recognition for deep neural networks as proof of this.
According to Managing Director Dr. Michael Förtsch, the chip technology available via a standard PCIe interface brings the power of photonics directly into real applications. For the first time, developers can explore the possibilities of the technology for complex, non-linear calculations. While a GPT-4 query today consumes ten times more power than a normal Internet search query, the photonic chips could reduce the energy consumption for this query by at least 30 times.
The LENA platform comprises Thin-Film Lithium Niobate (TFLN) on Insulator and enables precise light control at chip level. For example, a Fourier transformation, which requires millions of transistors in CMOS technology, can be performed with a single optical element.
